Learning What to Measure Next: Failure-Aware Closed-Loop Evidence Allocation for Protein Reliability

Protein design increasingly produces large sets of computationally plausible candidates, while physical measurements remain a scarce resource for assessing reliability. Using folding stability as a controlled reliability endpoint, we formulate screening as a sequential evidence-allocation problem and test whether a joint acquisition objective combining model disagreement with predicted low stability improves low-stability discovery under a fixed measurement budget. We instantiate this process as a closed-loop decision agent with explicit state, action, observation, measurement, update, and evidence-registry semantics. At each round the agent selects the next candidate batch and receives previously hidden experiment-derived folding-stability values through an offline measurement adapter. PRIB Failure-Aware combines bootstrap-ensemble disagreement with predicted low stability using within-pool percentile-rank aggregation. In a frozen retrospective replay over 30,000 MGnify Stability sequences, PRIB achieves higher low-stability recall than Random and generic uncertainty sampling at every tested checkpoint across five algorithmic seeds. At 8,192 measurements, recall is $53.95\pm0.77%$, compared with $27.62\pm0.81%$ for Random and $27.55\pm1.01%$ for Uncertainty. The results support objective-aligned experiment selection as a mechanism for allocating scarce physical evidence and define a traceable interface for extending the same decision loop to prospective laboratory tools.